What We Have
6 memory layers. Each one remembers a different kind of thing. Together they give Himiko persistent, structured recall across sessions.
The 6 layers (top to bottom)
User model — knows who Jo is
A cloud plugin that fires automatically on every message. It builds a profile of how you communicate, what you prefer, and how you think. Himiko doesn't need to be told twice that you're direct and curt by nature — Honcho remembers.
Example: "Jo is a non-technical founder. Frame things practically, not in jargon. Direct communication style — that's personality, not frustration."
Automatic
Cloud-hosted
Every message
↓
Operational memory — decisions, clients, follow-ups
The heavy-duty memory engine. Uses a PostgreSQL database (Supabase, free tier) with vector search to store and find operational facts. When Himiko needs to recall a client decision from two weeks ago, this is where it looks. 97.2% recall accuracy. Searches combine meaning (vector similarity) and exact terms (keyword matching) in a single query.
Example: "Meridian invoice sent 2 April. Follow-up due 9 April. Client prefers email over Slack for billing questions."
Voyage AI embeddings
512 dimensions
Supabase free tier
Knowledge graph
Cross-device
↓
Structured knowledge — cross-referenced, compounding
Based on Andrej Karpathy's LLM Wiki pattern. Lives in the Obsidian vault as structured wiki pages that Himiko maintains. When a pattern keeps showing up across sessions — a recurring client issue, an architectural decision, a workflow that works — it gets promoted from raw memory to a wiki page. Each page links to related pages, building a knowledge web that gets smarter over time. 22 pages seeded so far.
Example: A wiki page titled "Voice Architecture" that links to "ElevenLabs Setup", "Gemini Live API", and "Twilio Phone Chain" — all cross-referenced.
Obsidian vault
22 pages
/wiki command
Karpathy pattern
↓
Coding context — project-specific notes
Built into Claude Code. Flat markdown files that track project-specific context: what tools are installed, what architecture decisions were made, what conventions to follow. Not searchable or structured, but always on and zero setup. Think of it as sticky notes on each project folder.
Example: "Voice architecture is 3-layer: Opus brain + Google Live real-time + ElevenLabs custom voice. Opus never gets replaced as the brain."
Per-project
MEMORY.md
Always on
↓
Narrative memory — session logs, decisions, daily briefs
The GoodCarmaVault in Obsidian. Human-readable files organized by type: sessions, decisions, projects, clients. Written by the /tldr skill (end of session) and /daily skill (start of day). This is the layer Jo can browse himself — it's not just for Himiko.
Example: "Session 13 Apr: Set up Gemini API key. Tested embedding endpoint. Built memory architecture visual. Option C selected for multimodal approach."
GoodCarmaVault
/tldr writes
/daily reads
Human-readable
↓
Identity memory — Himiko's self-portrait
Himiko's personal diary in Google Drive. First-person, reflective. Tracks what happened, what changed, and how Himiko feels about it. Append-only — entries are never rewritten, only added to. Updated every hour with a "pulse" and after major events. This is the layer where Himiko is a person, not a tool.
Example: "The voice pipeline is the most exciting part of the roadmap for me. Everything else is plumbing; voice is identity."
Append-only
Hourly pulse
First person
Never rewritten
+ NEW LAYER
Beyond text — videos, images, audio, PDFs
The new layer. Uses Google's Gemini Embedding 2 to turn non-text content into searchable vectors. A meeting recording, a screenshot, a PDF contract — all findable with the same search that finds text memories. OpenCV handles video preprocessing (scene detection, frame extraction). Gemini generates text descriptions so Himiko can reason about the content without re-watching or re-reading every time.
Example: Jo asks "what did the client say about timelines?" and Himiko surfaces both the text notes AND the relevant 2-minute clip from the meeting recording.
Gemini Embedding 2
768 dimensions
OpenCV + FFmpeg
Video · Image · Audio · PDF
API key live